Hardware
Pixabay
Within 24 hours at the start of June, two of the most influential companies in computing placed the same bet. At GTC Taipei on June 1, Nvidia unveiled RTX Spark, a superchip it positioned as the foundation for a Windows PC rebuilt around personal AI agents. A day later at Build, Microsoft introduced Project Solara, a platform for devices designed to run agents in place of traditional applications. Both pitches rest on one premise that deserves scrutiny before any company rewrites its hardware roadmap. An agent, the argument goes, needs silicon and a device shaped for it.
For technology decision makers, the timing forces a question that was easy to defer a year ago. If agents become the way employees get work done, does that change what sits on the desk, in a pocket or in a badge clipped to a uniform. Or does the intelligence that matters stay in the cloud no matter what hardware a vendor ships. The answer shapes refresh cycles, device budgets and the assumptions behind every endpoint security plan.
RTX Spark pairs an Arm-based Grace CPU with a Blackwell graphics processor in a compact system carrying up to 128GB of unified memory, linked by Nvidia's NVLink interconnect. The design lets a large model run on the device without a round trip to a data center, and Nvidia says the platform can hold models of around 120 billion parameters with context windows reaching 1 million tokens. RTX Spark laptops and small desktops are due this fall from Asus, Dell, HP, Lenovo, Microsoft Surface and MSI. The framing throughout was the PC as a teammate rather than a tool.
Project Solara takes a different path to a similar idea. Rather than a chip, Microsoft showed a software platform built on a lightweight operating system derived from the Android Open Source Project, which it calls the Microsoft Device Ecosystem Platform. It ran on reference hardware using chips from Qualcomm and MediaTek, and Microsoft demonstrated two concept devices aimed at enterprise workers, a desk companion and a wearable badge. The platform leans on what Microsoft describes as just-in-time interfaces, where an agent generates the screen it needs for a task rather than forcing developers to redesign an app for every form factor. Early pilot partners include AccuWeather, Best Buy, CVS Health, Levi's and Target.
The contest over agent infrastructure has mostly played out in the data center, where Nvidia, Google with its Ironwood tensor processing units and AWS with Trainium compete to train and serve the models behind agentic systems. RTX Spark and Project Solara push that contest toward the edge, onto the devices people hold. Microsoft is not the only company chasing an agent platform, since Google, Salesforce and OpenAI are each building their own. It is among the first to argue that agents warrant a purpose-built device that is neither a phone nor a PC.
The two announcements also reveal a split in thinking. Nvidia wants the heavy computation to happen locally, which is why RTX Spark spends so much of its silicon on a capable GPU, while Microsoft built Solara around a chip-to-cloud model in which a low-power device handles input and security and the demanding inference runs in Azure. Nvidia treats the device as the engine, whereas Microsoft treats it as a secure front door to intelligence that lives elsewhere.
The strongest case against agent hardware came from outside either company. Writing at Stratechery, Ben Thompson argued that the ideal setup for a local agent is a strong CPU paired with calls to the cloud for inference, and that RTX Spark devotes die space to GPU cores that remain weaker than what a data center offers on memory and bandwidth. By that reading, a buyer pays a premium for local horsepower that the cloud still does better and cheaper.
Microsoft's own design hints at the same conclusion. If Solara devices summon cloud-hosted agents for the real work, the hardware on the desk is closer to a managed terminal than an AI workstation, which raises a fair question about how much specialized silicon a thin client truly needs. The timelines compound the doubt, because Solara is early by Microsoft's own admission, with a pilot now and broad availability not expected until late 2027 or 2028. Demand is also unproven, since most enterprise agents being deployed today in retail, finance and healthcare run well enough on the phones and laptops companies already own.
The useful move is to separate the orchestration and security layer from the inference layer, because the two announcements are strongest on the former and weakest on the latter. Local context, lower latency, data that never leaves the building and centralized device management are concrete benefits a CXO can test against real workflows. Running frontier models on a desk is a harder sell while the cloud keeps the lead on the largest models.
The practical step is to watch the pilots rather than the keynotes. If AccuWeather, CVS Health and Target report that purpose-built agent devices cut handling time or error rates in defined tasks, the category earns a place in the 2027 budget conversation. Until that evidence arrives, the safer read is that agents will reshape what people do with their existing hardware long before they justify buying something new.
此内容由惯性聚合(RSS阅读器)自动聚合整理，仅供阅读参考。 原文来自 — 版权归原作者所有。